Research On Pedestrian Data Augmentation Method Based On Correlation Guidance

With the increasing demand for public safety,a large number of intelligent monitoring devices are deployed in public places such as streets,campuses and communities.As an important research content of intelligent monitoring system,person re-identification is widely used in the field of target tracking and target retrieval.In recent years,with the rapid development of Deep Learning,data plays an essential role in person re-identification based on Deep Learning.Its scale and quality determine the accuracy of recognition and generalization of person re-identification.However,due to exist pedestrian datasets are limited in size,the high cost of manual labeling,the poor quality of pedestrian images generated by Generation Adversarial Network without relevant constraints,and the large difference of samples collected in different scenes,the accuracy of recognition and the ability of cross-domain recognition of person re-identification are greatly affected.Aiming at the above problems,this paper proposes a pedestrian Data Augmentation method based on distance correlation guidance and a pedestrian Data Augmentation method based on semantic correlation guidance.The specific contents are as follows.(1)Research on Person Data Augmentation in The DomainAiming at the problem that it is difficult to generate high-quality pedestrian images in the domain by Generation Adversarial Network without relevant constraints,a method of pedestrian Data Augmentation guided by distance correlation is proposed.Firstly,a local short-range correlation guidance mechanism is designed to improve the short-range dependence of image generation and the fine-grained feature expression ability of pedestrian image generation by fusing local multi-scale feature information.Secondly,a long-distance correlation guidance mechanism is designed to make more reasonable use of the global-feature distribution through the use of the External Attention mechanism to improve the overall visual quality of the generated pedestrian image.Finally,the three-network stability model is built by adding the adversarial re-discrimination network to the original structure of Generation Adversarial Network to improve the stability of network training.The practice results show that this method can effectively reduce the cost of manual labeling,improve the quality of pedestrian image generation,and promote the training of person re-identification.(2)Research on The Method of Cross-domain Person Data AugmentationAt present,the cross-domain use of the model is also an urgent problem in the field of person re-identification.Although the method of pedestrian Data Augmentation guided by distance correlation has been verified to be effective in image generation in the domain and can promote the training of the model,when it is applied to the cross-domain field of person re-identification,it is highly susceptible to interference from pedestrian image background and style factors,making it difficult to generate images with distinguishing features.Then the accuracy of person re-identification is seriously damaged.Therefore,in view of the limitations of the above methods,this paper proposes a semantic relevance guided pedestrian Data Augmentation as a supplement.Firstly,a multi-dimensional fine-grained semantic segmentation network is proposed to improve the accuracy of pedestrian image semantic analysis,avoid the interference of style and background factors,and provide guidance for subsequent pedestrian attribute generation.Secondly,designing a Generation Adversarial Network guided by semantic information,which is guided and restricted by semantic information to complete the generation of pedestrian corresponding attributes and the migration of style.The practice results show that this method effectively makes up for the shortcomings of the above method,and improves the cross-domain recognition ability of the person re-identification from the perspective of Data Augmentation.Finally,the research content of the article is summarized and analyzed.In view of the shortcomings in the research,this article analyzes and prospects it in the subsequent chapters.